Unsupervised feature learning based on autoencoder for epileptic seizures prediction

Epilepsy is one of the most common neurological diseases in the world. It’s essential to predict epileptic seizures since it can provide patients with enough time for timely treatment. Currently, electroencephalogram (EEG) analysis has been adopted as the most popular method of epileptic seizures prediction, of which one key element is extracting important EEG features. Conventional technologies of EEG analysis mostly utilize supervised learning methods with a mass of labeled data, which bring leakage risks to healthcare data. In addition, it’s difficult to achieve high accuracy of epileptic seizure prediction based on unsupervised learning methods with huge network parameters. Furthermore, the insufficiency of preictal data leads to overfitting challenges for deep learning algorithms. To deal with this problem, a data augmentation method based on randomly translation strategy is proposed to address the insufficient datasets without additional noise. In this paper, we propose an improved unsupervised feature learning method, residual convolution variational autoencoder with randomly translation strategy (RTS-RCVAE). Residual learning is embedded in the VAE model, which improves the model’s ability to converge in the unsupervised learning stage and reduces the loss of useful information. The proposed model is trained and verified via simulation using the public dataset CHB-MIT. The results indicate that the proposed model achieves a high accuracy rate of 98.43% and a false alarm rate of 0.009.